Intel fires 64-bit server salvo

By
iApplianceWeb
(06/24/02, 10:02:25 AM EDT)

Intel Corp. is mounting a fresh assault on the lucrative 64-bit server market in the next few weeks with the official introduction of its Itanium 2 processor and its first 16-way multiprocessing chip set.

Observers and analysts say it will take months, if not years, to establish the architecture, though many of them expect the company to make inroads over the long term, particularly against 64-bit server giants Sun Microsystems Inc. and, to a lesser extent, IBM Corp. This server battle of titans will play out largely on three fronts:

• Trade-offs in CPU integration and their impact on performance in complex multiprocessing systems. Here, the winners are far from clear.
• Server market credibility in generally software-based reliability features. Here, Sun and IBM have the advantage.
• And raw resources in design, manufacturing and marketing to wage a prolonged competitive fight. Here, Intel comes out ahead.

Advanced Micro Devices Inc. could act as a spoiler in this battle with its 64-bit extensions to the X86 architecture in its forthcoming Opteron processor line. Analysts praise the elegance and scalability of the Opteron architecture but say AMD may lack the system expertise and top hardware and software partners to gain broad market credibility.

The target is clear. A relatively small group of high-end multiprocessing systems - about 11 percent of all servers - command as much as 60 percent of all server revenue, according to International Data Corp. "In the last server out of 10, there's a lot of revenue," said Tom Garrison, director of enterprise platform marketing at Intel.

"If you are wondering about our strategy with Itanium, follow the money," said Lisa Hambrick, who heads enterprise processor marketing at Intel.

Intel will roll out Itanium 2 this summer at prices expected to range from $2,000 to $4,000. More than 20 OEMs - about half of them small regional outfits in Russia, China, Brazil and elsewhere - will ship 10 unique designs based on Itanium 2 by the end of the year, according to Vaughn Mackie, an Intel marketing manager.

Intel will seed the market for big multiprocessing systems with its E8870 chip set, its first to support 16-processor systems. Bull, NEC and Unisys will develop ASICs to field 32-way Itanium 2 systems.

Seven operating systems will be available by year's end, including two flavors of Windows, four Linux distributions and HP-UX. IBM's AIX, Sun's Solaris and Novell's Netware will not be available on Itanium 2.

At chip level, Intel initially claimed best performance on estimated base SPEC figures. However, IBM's Power4 architecture outshines Itanium 2 in integer figures. At the systems level, Intel claims a four-processor 1-GHz Itanium 2 system with 32 Gbytes of RAM will cost $41,000 and outperform a similarly configured $60,000 Sun V880 900-MHz Ultrasparc III system based on TPC-C benchmarks of the Transaction Processing Performance Council.

Cache cram


But Intel's approach is based on cramming lots of Level 3 cache on the processor: 3 Mbytes on Itanium 2, 6 Mbytes for the 2003 Madison version and more for a chip in 2004. The added cache gives oomph to a multiprocessor design based on four processors using a shared bus with separate memory controllers and switches linking four-processor nodes into larger systems.

That's inefficient, said Kevin Krewell, senior analyst with MDR/In-Stat. "You have to put 2 Mbytes or more of cache onboard to make the front-side-bus traffic levels acceptable for a shared bus scenario. AMD's architecture has better scalability. It's a more sophisticated architecture and requires less chips to build a system," he said.

Indeed, AMD builds a memory controller into its high-end Opteron line and can directly link up to four processors using onboard HyperTransport interfaces. AMD has yet to supply performance specs for its Opteron line, which will debut in dual-way versions late this year and four- and eight-way versions in 2003.

IBM has gone two steps further by building two PowerPC cores and a switching fabric onto its Power4 processor. However, both the memory controller and L3 cache are external.

That approach means Power4-based multiprocessors use the fewest chips and thus could cost the least to build and have the least system latency, said Joel Tendler, an IBM program director for technology assessment. Specifically, signals acquire a maximum of 15 percent latency on their farthest routes in Power4 systems compared to a 200 percent latency in most shared bus scenarios, he said.

IBM claims its 32-way p690 Regatta systems, its first Power4 servers launched last October, hit 403 TPC-C transactions/minute at a cost of $17.85 per transaction. That figure bested 64-way Hewlett-Packard systems, he said. IBM has sold 1,000 of the systems as of June 1 and is rolling the Power4 into multiple new server systems.

IBM will move the 0.18-micron Power4 into a 0.13-micron process next year to boost speed above the current 1.3-GHz rating and reduce power consumption. A Power5 version in 2004 will support 64-way (32 dual-core chip) systems.

Dual designs


Nevertheless, IBM is hedging its bets, designing ASICs for both high-end Itanium and Power4 systems. "That's an advantage if it doesn't confuse its sales staff and customers," said analyst Krewell.

For its part, "Sun is lagging in performance," and it is keeping mum on its road map, Krewell said. But "Sun is still the incumbent here," and neither performance nor price is the top issue for server buyers. "Reliability is," he said.

Both Sun and IBM shine here. IBM, for instance, already offers software features to make a 32-way Power4 server look to the user like 16 separately addressable systems. Future generations will expand on such capabilities.

With its future Ultrasparc V processor, Sun will stake out a unique position between the throughput focus of IBM's Power4 and the single-stream focus of Itanium, said David Yen, vice president and general manager of Sun's processor products group. Yen earlier said the 0.09-micron Ultrasparc V, due out "in a couple years," will flexibly handle both thread- and instruction-level parallelism championed separately by IBM and Intel today.

Integration of more system-level capabilities will be another key differentiator in future microprocessors, Yen said in a phone interview this past week. But Yen would not reveal more details about the Ultrasparc V or Ultrasparc IV, a 0.13-micron enhancement and shrink of the existing Ultrasparc III due out at year's end.

Yen did take a swipe at Itanium as a brute force effort to create a general-purpose server CPU. "At 421 mm² and 130 watts, Itanium is more of a lab prototype than a volume processor. They are trying to buy time. They can't afford to let software developers lose enthusiasm," he said.

In that arena, AMD trails even Intel in availability of server-class software for core reliability features. That's because so far the Opteron family lacks top tier OEM and software backers.

Size counts


"They need to get a big vendor behind them. You can't crack into that space with tier 3 suppliers," said Krewell, who, like other analysts, suggest Sun could be that key partner. Without such backing Opteron could be limited to annual unit sales of a few hundred thousand in largely two-way servers.

Intel clout comes from its deep pockets to endure what many say will be a 10-year battle with Sun and IBM. Indeed, the company has five follow-on processors to the Itanium 2 in the works, including two in 2003, one in 2004 and multicore and multithreaded CPUs for delivery beyond that.

"Intel is turning the crank on these things and will move things along faster than Sun with new chips every year. If you are competing with Intel in manufacturing it's pretty scary," said Krewell.

"Intel's volume will drive the process technology and I think that will make them the hands-down leader in the long run," said Dave Houseman, chief technologist of Unisys Corp.'s Systems and Technologies Division (Blue Bell, Pa.), which will offer 32-way Itanium 2 systems this year using some of its own ASICs.

That doesn't make Intel a shoe-in. Analysts and observers say it will take nine months or more just to deliver a fully production-ready stack of operating systems and applications on Itanium 2 to run app benchmarks.

"I don't think the battle will be fully engaged until early 2003. It will take a while for the dust to settle," said Krewell.

Among other challenges for Itanium 2, the IT market is expected to see flat to low growth this year and next. Also, it is generally averse to new technologies. Some say Intel faces huge marketing problems because most potential users are not even aware of Itanium. And some OEMs are still very cautious about backing the technology given the market failure of the underpowered first-generation Itanium, a.k.a. Merced.

"We're in a wait-and-see posture, trying to understand what the market wants," said Jim Pike, a director of server architecture and technology at Dell Computer and formerly the designer of Intel's eight-way chip set for the X86.

"The market certainly did not commercially accept Merced. You could say that was because Intel preannounced it to death, or you could say it was a wholesale resistance to Intel's 64-bit story," he said.

Some analysts suggest OEMs like Dell would rather Intel deliver a 64-bit version of its X86 Xeon server chips, reportedly in the works under the code name Yamhill. That could be a simpler product to deliver and promise broader sales for a high-volume manufacturer like Dell at least in the short term, they say.

But that approach would miss the target, said Intel's Hambrick. "You can't hit this space with extensions to the existing architecture. You need to go to a 64-bit flat memory space and you need bigger page sizes," she said.